Headlines Since

Catch up on the biggest news stories that happened since a date you choose. After being away from the news (e.g. on vacation), get a small curated list of the most significant individual stories from your chosen start date to today.

The code in this project was bootstrapped by Cohere's north-mini-code-1.0 with OpenCode and finished with Anthropic's Claude Code.

Goals

• Biggest individual stories: Ranked by significance (public interest, number of sources, soft authority), not by how recent they are. Reports of the same event are merged, and a gentle penalty on repeated topics keeps one big ongoing story (e.g. a war) from filling the whole list, without hiding its distinct developments.
• Elm frontend: Clean, simple UI and all app logic written in Elm. The only JavaScript is src/index.js, which starts the Elm app.
• Client-side data: Fetches news in the browser, from Wikipedia directly and from RSS feeds via CORS proxies.
• Pure static: Deployed as static HTML/CSS/JS to GitHub Pages (no backend).

Data Sources

• Wikipedia's Current Events portal: an editor-curated list of each day's notable events, with citations. One page per day, so it covers any start date. Fetched directly from the Wikipedia API (no proxy needed).

• 23 outlets' RSS (2.0 and 1.0) and Atom feeds, fetched through CORS proxies (corsproxy.io if an API key is configured, falling back to allorigins.win), except DW, The New York Times and The New Yorker, whose feeds can be fetched directly:

  • Paged back to the start date (up to 10 pages each): Capital B, Common Dreams, Mother Jones, Prism, The 19th, The Intercept, The Nation, Truthdig.
  • Latest items only (their feeds can't be paged; they reach back from about a day to two months): Axios, Democracy Now!, DW, Jacobin, NPR, Politico, Popular Information, ProPublica, The Atlantic, The Guardian (US), The Marshall Project, The New York Times, The New Yorker, Truthout, Vox.

Outlets' articles can't be ranked against Wikipedia events (they have no Wikipedia topic, so no page views). An article is attached to a ranked story as a related report when it clearly reports the same event; the rest are listed per outlet below the ranked stories, newest first. See Matching outlets' articles.

The ranked stories appear as soon as Wikipedia has loaded; each outlet's section fills in when that outlet's feed has loaded.

Article links open the newest archived copy on archive.today (https://archive.ph/newest/<article URL>), so paywalled articles can be read; hovering a link shows the original address. If an article hasn't been archived yet, archive.ph shows a "No results" page with a button to archive it there and then. Tracking parameters (e.g. utm_source, the BBC's at_medium, DW's maca) are removed first, because archived copies are usually of the plain address.

Removed sources:

• Google News RSS: Google answers requests from CORS proxies with a "Sorry..." block page (HTTP 503).
• LA Times RSS: refuses requests from corsproxy.io (HTTP 403), and allorigins.win hangs until it times out.
• BBC News RSS: its headlines link to no Wikipedia article, so they can't be given a public-interest score or reliably matched to Wikipedia events about the same story (tested: exact citation URLs matched 0 of 33 headlines; title word overlap and Wikipedia search were both wrong too often). Its significant stories were already covered by Wikipedia.

Configuration

The corsproxy.io API key is read from the VITE_CORSPROXY_API_KEY environment variable at build time and passed into Elm as a flag. It is optional; without it, only allorigins.win is used.

For local development, copy the example file and fill in your key:

cp .env.example .env

For CI deployment, set the same name, VITE_CORSPROXY_API_KEY, as a GitHub repository secret; the deploy workflow passes it to the build as an environment variable. corsproxy.io also needs the deployed site's domain (alxndr.github.io) allowed in its dashboard.

The key ends up in the public JavaScript bundle, as any key used from the browser must; corsproxy.io's domain allowlist is what stops others from using it.

Development

Quick start with Vite

Vite with vite-plugin-elm provides hot module reloading for Elm.

# Install dependencies
npm install

# Optional: configure the corsproxy.io API key
cp .env.example .env

# Start dev server
npm run dev

Visit http://localhost:5173.

Build for production

npm run build

The built site will be in the build/ directory. Preview it locally with:

npm run preview

Production builds use the /headlines-since/ base path (for GitHub Pages), so the preview is at http://localhost:4173/headlines-since/.

Tests and formatting

npm test              # elm-test, in tests/
npm run format        # elm-format, rewrites files
npm run format:check  # elm-format, fails if anything needs formatting

The Tests workflow (.github/workflows/test.yml) runs the format check, tests, and a build. It runs on pull requests, and on pushes to main as the first step of the deploy workflow, which doesn't deploy unless it passes.

Using mise for tool versions

With mise, the correct Elm version will be used automatically (pinned in .tool-versions).

mise install
mise exec -- npm run dev

Deployment

GitHub Actions workflow (.github/workflows/deploy.yml) runs the Tests workflow, then builds the app with Vite and deploys to GitHub Pages, on push to main. If the format check, tests or build fail, nothing is deployed. The build step reads the corsproxy.io key from the VITE_CORSPROXY_API_KEY secret. Enable Pages in repo settings (source: GitHub Actions).

Code layout

Module	Purpose
src/Main.elm	The form and results; starts the Wikipedia ranking and each outlet's feed loading in parallel, and attaches outlets' articles to ranked stories
src/Feeds.elm	Fetches Wikipedia's Current Events, and outlets' feeds (paging back where possible; most via the CORS proxies)
src/Rss.elm	Parses RSS and Atom feeds into stories
src/WikipediaCurrentEvents.elm	Parses Wikipedia Current Events day pages into stories
src/Rfc822.elm	Parses RSS <pubDate> dates
src/PageViews.elm	Fetches Wikipedia page views for story topics
src/SameEvent.elm	Recognises different reports of the same event
src/OutletMatch.elm	Matches outlets' articles to Wikipedia events
src/Rank.elm	Scores and picks the top stories
src/Story.elm	The Story type
src/ArchiveLink.elm	Turns article links into archive.ph links

Algorithm

Only Wikipedia events are ranked. They're filtered to those published from the start date through today (in your time zone), then scored as a weighted sum of:

• Public interest (0.5): average daily Wikipedia page views during the date range for the event's most specific topic (e.g. for an event filed under "2026 Iran war › 2026 Iran war fuel crisis", the fuel crisis article), on a log scale. Events without a topic heading (about a fifth of them) get no public-interest score. For date ranges starting more than 60 days ago, only the 50 most-cited topics are looked up, because each needs its own request.
• Sources (0.15): how many news reports the event cites
• Authority (0.1): whether the first outlet the event cites is a major one

Sports and arts stories get half the score, because they draw far more page views than their significance warrants.

Two things are deliberately not factors:

• Recency: when catching up after time away, a big story from the first day matters as much as one from today.
• Keyword lists (e.g. "killed", "election"): page views measure significance more directly, and a hand-written list skews toward whatever kinds of news it happens to cover.

Stories are then picked one at a time, highest score first:

• Same event, merged: a report of an event that's already been picked (e.g. a disaster's death toll updated on later days) is listed under that story as a related report, rather than taking up a slot. Two Wikipedia events count as the same event if they're at most 3 days apart and link to largely the same Wikipedia articles (weighted Jaccard similarity of 0.55 or more, where articles linked from many events, like "United States", count for little).
• Repeated topics, penalised: each further story from a topic that's already been picked (e.g. a third story filed under "2026 Iran war") scores 15% less than the one before, so other news can compete.

Duplicate URLs are only considered once.

Limitations

• The most recent day is thin. Wikipedia's page for a day fills in as the day goes on (e.g. at 01:39 UTC on 3 October 2026, that day's page had no events yet, 2 October had 14 and 1 October had 22), so the ranked list has little from today. Outlets' articles from today still appear in their sections, unranked.
• Long date ranges are partly covered, and slower. Paged outlets' feeds go back at most 10 pages each (e.g. about 2 weeks for Mother Jones, 3 months for The Nation and The Intercept, 6 months for Common Dreams), loaded one page after another (a month-long range took about 26 seconds for every outlet to load, though the ranked stories appeared in 4), the other outlets only have their latest items (a day to two months), and for ranges starting more than 60 days ago, page views are looked up for only the 50 most-cited topics. Each outlet's section notes when it couldn't reach back to the start date.
• Proxy quota. corsproxy.io's free plan allows 10,000 requests a month, and each outlet feed page is one request: about 25 per search for a range of a week or so, about 47 for a month, and at most 92 (8 paged outlets at 10 pages each, plus 12 other proxied outlets).
• Many articles haven't been archived yet. In a sample of 14 linked articles, 5 already had an archived copy on archive.ph; for the rest, readers land on archive.ph's "No results" page and have to ask it to archive the article (one click, then a short wait).
• Outlets' articles rarely match. See below. Straight-news outlets (e.g. DW, The New York Times, Axios) match more often than opinion and investigative ones, since they report the same events Wikipedia lists.

Matching outlets' articles

An outlet's article (title plus its RSS summary) is compared with each Wikipedia event published within 3 days of it, as sets of words weighted by how rare each word is among the events (cosine similarity of TF-IDF-style weights). It's attached to the event if the similarity is at least 0.25.

This is deliberately strict. Tested on 142 real headlines from the BBC, The Nation, Common Dreams and Mother Jones, hand-checked against the Wikipedia events: only 12 had a matching event at all (these outlets mostly publish opinion and investigations, which Wikipedia doesn't list as events). At 0.25 the matcher found 2 of the 12 and made no wrong matches; looser settings found more but quickly made wrong ones, such as matching "Revolution or No Revolution" to an unrelated arrest in Myanmar.